1. CALTECH CS137 Winter2006 -- DeHon CS137: Electronic Design Automation Day 4: Jan 18, 2006 Concept Generation

2. CALTECH CS137 Winter2006 -- DeHon Today Specifications/requirements/goals Concept Generation

3. CALTECH CS137 Winter2006 -- DeHon Specification Before can start solving –Need to know what your solving Before can evaluate “goodness” –Need to know what trying to accomplish –Priorities

4. CALTECH CS137 Winter2006 -- DeHon Quantitative Specification Typically need to –turn vague specification –into something quantifiable/measurable Something concrete

5. CALTECH CS137 Winter2006 -- DeHon Specifications More of a trick/challenge for things that don’t have natural metrics –From: Door closes easily –To: Door latches with <5 lbs. of force… –From: reliable –To: 1 failure in 10 9 hours of operation 1 undetected failure …

6. CALTECH CS137 Winter2006 -- DeHon Our Specifications [talk about]

7. CALTECH CS137 Winter2006 -- DeHon Establish Priorities Which requirement is king? Which have flexibility? –Can tradeoff?

8. CALTECH CS137 Winter2006 -- DeHon Establish Targets Acceptable Ranges Ideal Usually Relative to alternatives/competitors

9. CALTECH CS137 Winter2006 -- DeHon Our Targets and Priorities [sketch] –Delay in Solution –Area of logic in Solution –Energy of Solution –Reliability of Solution –Runtime of algorithm –Area for hardware engine to solve problem Your problem formulation should capture priorities

10. CALTECH CS137 Winter2006 -- DeHon Concept Generation

11. CALTECH CS137 Winter2006 -- DeHon Goal of Concept Generation Fully explore design space –Not miss options –Open minds to possibilities –Know what can do Competitors can do

12. CALTECH CS137 Winter2006 -- DeHon Components (Outline) Structured/Systematic Approach Partial Solutions Uncertainty/Unknown Documentation

13. CALTECH CS137 Winter2006 -- DeHon Non-Atomic Insight Nothing will ever be attempted, if all possible objections must be first overcome. --- Samuel Johnson, 1759.

14. CALTECH CS137 Winter2006 -- DeHon Uncertainty “Do you remember what I have taught you? Most intelligent beings prefer to live in certainty than uncertainty. Rather than accept uncertainty, they will discount the input of their own senses. It is through this mechanism that mages manipulate the perception of others.” --- Elric to Galen upon his [Elric's] death Invoking the Darkness, Jeanne Cavelos

15. CALTECH CS137 Winter2006 -- DeHon (This slide intentionally left [almost] blank in this copy)

16. CALTECH CS137 Winter2006 -- DeHon Concept Generation: Steps Start refined problem specification 1.Clarify/decompose 2.Search –Externally –Internally / brainstorm 3.Explore Systematically 4.Reflect

17. CALTECH CS137 Winter2006 -- DeHon Decompose Separate components must solve –Divide and conquer Maybe attack w/ separate passes through process Focus on bottleneck –Critical path/subproblem –Where most risk is Most innovation needed

18. CALTECH CS137 Winter2006 -- DeHon Search Externally Literature (standing on the shoulders of giants) –Published / library / IBID –Web –Patents –Pattern Catalog Customers Experts Benchmark related / competitive  Understand state-of-art

19. CALTECH CS137 Winter2006 -- DeHon Search Internally/Brainstorm Common bugs: –Reject things too early “Nothing…” quote –Latch onto few ideas Get stuck in own local minima

20. CALTECH CS137 Winter2006 -- DeHon Search Internally/Brainstorm Different mode of thinking –“Pattern” of thinking Generate lots of ideas Suspend judgment / no squashing Barr: grow mode / acid mode Find good pieces, incomplete ideas Wishful thinking Comfortable with uncertainty

21. CALTECH CS137 Winter2006 -- DeHon Brainstorm Throw out ideas Write them down … draw pictures De-personalize –Name by characteristics not person –Not attacking person/idea –Group ownership –Everyone contribute to debugging Mix-and-match Good science –Find best ideas, avoid NIH

22. CALTECH CS137 Winter2006 -- DeHon People and Brainstorming More brains generally good –Maybe limit size of group at a time –Different perspectives good Group ownership of final product –Involve stake holders –Build confidence in conclusion –Get buy in

23. CALTECH CS137 Winter2006 -- DeHon Systematic Exploration Parameterized Design Space Concept combination matrix Classification tree Architecture/Pattern Tree Goals: –Find holes –Over-emphasis –Make sure think about all cross products..

24. CALTECH CS137 Winter2006 -- DeHon Design Explore for SMVM Transactional SDF GraphStep Organize parallelism? Too big?/Scale? Operator Sharing Communication?Store Data? Complex Op? Topology?Style? Ring Time Mux Local Mem VLIW Compute Managed? C-Slow Cycles? Pipeline

25. CALTECH CS137 Winter2006 -- DeHon Reflect …really throughout Comfortable solving problem? –Specification adequate? Process

26. CALTECH CS137 Winter2006 -- DeHon Time Permitting

27. CALTECH CS137 Winter2006 -- DeHon Design Examples SMVM

28. CALTECH CS137 Winter2006 -- DeHon Problem Compute: x=Ab Iterative: b i =x i-1

29. CALTECH CS137 Winter2006 -- DeHon Parallelism? Compute: x=Ab 1.Dot products independent 2.Multiplies within dot product independent 3.Sum of dot products associative? Maybe not for floating point

30. CALTECH CS137 Winter2006 -- DeHon (Special) Problem Characteristics Matricies can be large –N×N with N in thousands Matrix A is sparse –E.g. < 100 non-zero entries in a row

31. CALTECH CS137 Winter2006 -- DeHon Compressed Sparse Row (CSR) matrix representation 1.Remove zeros 2.Concatenate rows 3.Store column indices 4.Store row starting positions Sequential algorithm: Compute dot products top to bottom Iterate over matrix_value, column_index

32. CALTECH CS137 Winter2006 -- DeHon What Architecture is Appropriate? Accept DP FP not assoc. –Serialize dot products –  BSP>Data Parallel>GraphStep (SDF) 2 1 0 4 3

33. CALTECH CS137 Winter2006 -- DeHon Design Explore Transactional SDF GraphStep Organize parallelism? Too big?/Scale? Coarse-Grained Time Multiplexing Fine-Grained Time Multiplexing Common Element Sharing for Regular Graphs Common Operator Sharing for General Graphs Scheduled Operator Sharing Modulo Scheduling Multicontext (physical) Datapath Sizing and Serialization Synthesis Objective Function Tradeoffs Function Unit Binding Sequential vs. Parallel Implementation

34. CALTECH CS137 Winter2006 -- DeHon Design Explore Transactional SDF GraphStep Organize parallelism? Too big?/Scale? Coarse-Grained Time Multiplexing Fine-Grained Time Multiplexing Common Element Sharing for Regular Graphs Common Operator Sharing for General Graphs –Homogeneous FP MPY/ADD Scheduled Operator Sharing Modulo Scheduling Multicontext (physical) Datapath Sizing and Serialization Synthesis Objective Function Tradeoffs Function Unit Binding Sequential vs. Parallel Implementation

35. CALTECH CS137 Winter2006 -- DeHon Scaling? Accept DP FP not assoc. –Serialize dot products –Share operator for multiple dot products –Scaling: number of dot products/PE 2 1 0 4 3

36. CALTECH CS137 Winter2006 -- DeHon Design Explore Transactional SDF GraphStep Organize parallelism? Too big?/Scale? Operator Sharing Communication?Store Data?Compute Managed?

37. CALTECH CS137 Winter2006 -- DeHon Design Explore Transactional SDF GraphStep Organize parallelism? Too big?/Scale? Operator Sharing Communication?Store Data?Compute Managed? Shared vs. Distributed Memory On-Chip vs. Off-chip? …  on chip block rams

38. CALTECH CS137 Winter2006 -- DeHon Design Explore Transactional SDF GraphStep Organize parallelism? Too big?/Scale? Operator Sharing Communication?Store Data?Compute Managed? Topology?Style? Mesh Tree Ring/Bus Configured Packet Switch Time Mux

39. CALTECH CS137 Winter2006 -- DeHon Design Explore Transactional SDF GraphStep Organize parallelism? Too big?/Scale? Operator Sharing Communication?Store Data?Compute Managed? Revisit Arch. Decisions Dataflow Sequenced VLIW

40. CALTECH CS137 Winter2006 -- DeHon Design Explore Transactional SDF GraphStep Organize parallelism? Too big?/Scale? Operator Sharing Communication?Store Data? Complex Op? Topology?Style? Ring Time Mux Local Mem VLIW Compute Managed? C-Slow Cycles? Pipeline

41. CALTECH CS137 Winter2006 -- DeHon Discuss Assumptions Refine/elaborate specification Decompose problem

42. CALTECH CS137 Winter2006 -- DeHon Admin Friday –No class –Problem Formulation Due Monday: brainstorming session Wednesday: Nachiket on GraphStep programming